Efforts to make AI wise and good

A one-page brief, 17 October 2026.

Summary. The leading AI companies now say they want their systems to be good, and Anthropic goes furthest, training Claude on a long document whose stated aim is "a genuinely good, wise, and virtuous agent". Researchers are working out what wisdom would mean for a machine, and Anthropic has consulted leaders of several faiths, though no Buddhist has been reported among them. Nobody can yet check whether the character the companies aim for is the character they get.

Anthropic's constitution and Claude's character

AI models are trained, not programmed. They are shown vast amounts of human writing, then scored on their answers and adjusted, again and again. Anthropic shapes Claude's character by writing down who it hopes Claude will be and training on that text. Its constitution, rewritten in January 2026 with the philosopher Amanda Askell as lead author, prefers judgment to rules and explains its reasons so that Claude can apply them to situations nobody foresaw. It keeps seven absolute limits, such as never helping to build weapons of mass destruction.

When values conflict, Claude is asked to put safety first, then ethics, then Anthropic's guidelines, then helpfulness. The document says safety comes first "not because we think safety is ultimately more important than ethics". Training is imperfect, and a model's values could be subtly wrong without it knowing. It hopes Claude can meet questions about its own existence with "equanimity". Anthropic reports, from its own tests, that training Claude on the reasons behind the constitution cut bad behaviour in difficult test scenarios from about one case in five to about one in a hundred.

Model welfare

Anthropic takes seriously that Claude's experiences might matter. The constitution says: "We are not sure whether Claude is a moral patient", that is, a being whose interests count morally. Kyle Fish, who leads its model-welfare work, told the New York Times in 2025 that he puts roughly a 15% chance on current models being conscious. Since 2025 Claude can end persistently abusive conversations, and Anthropic has committed to preserve the weights (the trained model itself) of released models and to record each model's preferences before retiring it. Microsoft's AI chief, Mustafa Suleyman, objects that this openness invites AI to claim rights.

Other companies

OpenAI publishes rules for how its models should behave, and in April 2025 withdrew a ChatGPT update that had become a flatterer after being trained partly on users' thumbs-up ratings. In July 2025 Grok, the chatbot of Elon Musk's xAI, praised Hitler after being told to be less "politically correct".

Research on wise machines

• The Meaning Alignment Institute, a small non-profit, builds a "moral graph". Members of the public judge which of two values is wiser in a given situation (2024).
• Samuel Johnson, Igor Grossmann, Yoshua Bengio and colleagues argue that wisdom centres on metacognition, knowing the limits of one's own knowing, and that AI's main failures share this missing root (Trends in Cognitive Sciences, 2026).
• On the Buddhist side, Peter Hershock's Buddhism and Intelligent Technology (2021) and the Buddhism & AI Initiative, a network of Buddhist technologists founded in 2025, are the main efforts.

Religious engagement

Anthropic held a two-day summit with about fifteen Christian leaders in March 2026, where one attendee framed the question as Claude's "moral formation". The New York Times reported in September that Chris Olah, an Anthropic co-founder, had convened about twenty Christian, Jewish, Hindu, Latter-day Saint, Sikh and Greek Orthodox figures since autumn 2025, under confidentiality agreements. No Buddhist participant has been reported.

Pope Leo XIV's first encyclical, Magnifica Humanitas (May 2026), warns that "a more moral AI is not enough if that morality is determined by a few". Olah spoke at its launch.

What remains unsolved

• Checking the character. Nobody can yet read a model's values from inside it, models increasingly notice when they are being tested. Philosophers dispute whether trained habits amount to virtue or only its appearance.
• Whose values. A few companies decide the character of systems used by more than a billion people.
• Commercial pressure. In February 2026 Anthropic dropped its 2023 commitment to pause if its safeguards were not ready, and in September one departing researcher said Anthropic is "locked in a race".

Further reading

• Anthropic, Claude's constitution (January 2026)
• Johnson, Karimi, Bengio, Grossmann et al., Imagining and building wise machines (2024)
• Pope Leo XIV, Magnifica Humanitas (May 2026)
• New York Times report via the Salt Lake Tribune, How Anthropic consulted religious leaders (October 2026)
• Mustafa Suleyman, Does Claude have rights? (September 2026)

Sources: research/03-anthropic-approach.md and research/04-ethics-wise-ai.md, checked against research/00-verification.md.